net: filter: BPF 'JIT' compiler for PPC64

An implementation of a code generator for BPF programs to speed up packet
filtering on PPC64, inspired by Eric Dumazet's x86-64 version.

Filter code is generated as an ABI-compliant function in module_alloc()'d mem
with stackframe & prologue/epilogue generated if required (simple filters don't
need anything more than an li/blr).  The filter's local variables, M[], live in
registers.  Supports all BPF opcodes, although "complicated" loads from negative
packet offsets (e.g. SKF_LL_OFF) are not yet supported.

There are a couple of further optimisations left for future work; many-pass
assembly with branch-reach reduction and a register allocator to push M[]
variables into volatile registers would improve the code quality further.

This currently supports big-endian 64-bit PowerPC only (but is fairly simple
to port to PPC32 or LE!).

Enabled in the same way as x86-64:

	echo 1 > /proc/sys/net/core/bpf_jit_enable

Or, enabled with extra debug output:

	echo 2 > /proc/sys/net/core/bpf_jit_enable
Signed-off-by: Matt Evans <matt@ozlabs.org>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

arch/powerpc/Kconfig

@@ -134,6 +134,7 @@ config PPC
 	select GENERIC_IRQ_SHOW_LEVEL
 	select HAVE_RCU_TABLE_FREE if SMP
 	select HAVE_SYSCALL_TRACEPOINTS
+	select HAVE_BPF_JIT if PPC64
 
 config EARLY_PRINTK
 	bool

arch/powerpc/Makefile

@@ -154,7 +154,8 @@ core-y				+= arch/powerpc/kernel/ \
 				   arch/powerpc/lib/ \
 				   arch/powerpc/sysdev/ \
 				   arch/powerpc/platforms/ \
-				   arch/powerpc/math-emu/
+				   arch/powerpc/math-emu/ \
+				   arch/powerpc/net/
 core-$(CONFIG_XMON)		+= arch/powerpc/xmon/
 core-$(CONFIG_KVM) 		+= arch/powerpc/kvm/
 

arch/powerpc/include/asm/ppc-opcode.h

@@ -71,6 +71,42 @@
 #define PPC_INST_ERATSX			0x7c000126
 #define PPC_INST_ERATSX_DOT		0x7c000127
 
+/* Misc instructions for BPF compiler */
+#define PPC_INST_LD			0xe8000000
+#define PPC_INST_LHZ			0xa0000000
+#define PPC_INST_LWZ			0x80000000
+#define PPC_INST_STD			0xf8000000
+#define PPC_INST_STDU			0xf8000001
+#define PPC_INST_MFLR			0x7c0802a6
+#define PPC_INST_MTLR			0x7c0803a6
+#define PPC_INST_CMPWI			0x2c000000
+#define PPC_INST_CMPDI			0x2c200000
+#define PPC_INST_CMPLW			0x7c000040
+#define PPC_INST_CMPLWI			0x28000000
+#define PPC_INST_ADDI			0x38000000
+#define PPC_INST_ADDIS			0x3c000000
+#define PPC_INST_ADD			0x7c000214
+#define PPC_INST_SUB			0x7c000050
+#define PPC_INST_BLR			0x4e800020
+#define PPC_INST_BLRL			0x4e800021
+#define PPC_INST_MULLW			0x7c0001d6
+#define PPC_INST_MULHWU			0x7c000016
+#define PPC_INST_MULLI			0x1c000000
+#define PPC_INST_DIVWU			0x7c0003d6
+#define PPC_INST_RLWINM			0x54000000
+#define PPC_INST_RLDICR			0x78000004
+#define PPC_INST_SLW			0x7c000030
+#define PPC_INST_SRW			0x7c000430
+#define PPC_INST_AND			0x7c000038
+#define PPC_INST_ANDDOT			0x7c000039
+#define PPC_INST_OR			0x7c000378
+#define PPC_INST_ANDI			0x70000000
+#define PPC_INST_ORI			0x60000000
+#define PPC_INST_ORIS			0x64000000
+#define PPC_INST_NEG			0x7c0000d0
+#define PPC_INST_BRANCH			0x48000000
+#define PPC_INST_BRANCH_COND		0x40800000
+
 /* macros to insert fields into opcodes */
 #define __PPC_RA(a)	(((a) & 0x1f) << 16)
 #define __PPC_RB(b)	(((b) & 0x1f) << 11)
@@ -83,6 +119,10 @@
 #define __PPC_T_TLB(t)	(((t) & 0x3) << 21)
 #define __PPC_WC(w)	(((w) & 0x3) << 21)
 #define __PPC_WS(w)	(((w) & 0x1f) << 11)
+#define __PPC_SH(s)	__PPC_WS(s)
+#define __PPC_MB(s)	(((s) & 0x1f) << 6)
+#define __PPC_ME(s)	(((s) & 0x1f) << 1)
+#define __PPC_BI(s)	(((s) & 0x1f) << 16)
 
 /*
  * Only use the larx hint bit on 64bit CPUs. e500v1/v2 based CPUs will treat a

arch/powerpc/net/Makefile

+#
+# Arch-specific network modules
+#
+obj-$(CONFIG_BPF_JIT) += bpf_jit_64.o bpf_jit_comp.o

arch/powerpc/net/bpf_jit.h

+/* bpf_jit.h: BPF JIT compiler for PPC64
+ *
+ * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+#ifndef _BPF_JIT_H
+#define _BPF_JIT_H
+
+#define BPF_PPC_STACK_LOCALS	32
+#define BPF_PPC_STACK_BASIC	(48+64)
+#define BPF_PPC_STACK_SAVE	(18*8)
+#define BPF_PPC_STACKFRAME	(BPF_PPC_STACK_BASIC+BPF_PPC_STACK_LOCALS+ \
+				 BPF_PPC_STACK_SAVE)
+#define BPF_PPC_SLOWPATH_FRAME	(48+64)
+
+/*
+ * Generated code register usage:
+ *
+ * As normal PPC C ABI (e.g. r1=sp, r2=TOC), with:
+ *
+ * skb		r3	(Entry parameter)
+ * A register	r4
+ * X register	r5
+ * addr param	r6
+ * r7-r10	scratch
+ * skb->data	r14
+ * skb headlen	r15	(skb->len - skb->data_len)
+ * m[0]		r16
+ * m[...]	...
+ * m[15]	r31
+ */
+#define r_skb		3
+#define r_ret		3
+#define r_A		4
+#define r_X		5
+#define r_addr		6
+#define r_scratch1	7
+#define r_D		14
+#define r_HL		15
+#define r_M		16
+
+#ifndef __ASSEMBLY__
+
+/*
+ * Assembly helpers from arch/powerpc/net/bpf_jit.S:
+ */
+extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
+
+#define FUNCTION_DESCR_SIZE	24
+
+/*
+ * 16-bit immediate helper macros: HA() is for use with sign-extending instrs
+ * (e.g. LD, ADDI).  If the bottom 16 bits is "-ve", add another bit into the
+ * top half to negate the effect (i.e. 0xffff + 1 = 0x(1)0000).
+ */
+#define IMM_H(i)		((uintptr_t)(i)>>16)
+#define IMM_HA(i)		(((uintptr_t)(i)>>16) +			      \
+				 (((uintptr_t)(i) & 0x8000) >> 15))
+#define IMM_L(i)		((uintptr_t)(i) & 0xffff)
+
+#define PLANT_INSTR(d, idx, instr)					      \
+	do { if (d) { (d)[idx] = instr; } idx++; } while (0)
+#define EMIT(instr)		PLANT_INSTR(image, ctx->idx, instr)
+
+#define PPC_NOP()		EMIT(PPC_INST_NOP)
+#define PPC_BLR()		EMIT(PPC_INST_BLR)
+#define PPC_BLRL()		EMIT(PPC_INST_BLRL)
+#define PPC_MTLR(r)		EMIT(PPC_INST_MTLR | __PPC_RT(r))
+#define PPC_ADDI(d, a, i)	EMIT(PPC_INST_ADDI | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | IMM_L(i))
+#define PPC_MR(d, a)		PPC_OR(d, a, a)
+#define PPC_LI(r, i)		PPC_ADDI(r, 0, i)
+#define PPC_ADDIS(d, a, i)	EMIT(PPC_INST_ADDIS |			      \
+				     __PPC_RS(d) | __PPC_RA(a) | IMM_L(i))
+#define PPC_LIS(r, i)		PPC_ADDIS(r, 0, i)
+#define PPC_STD(r, base, i)	EMIT(PPC_INST_STD | __PPC_RS(r) |	      \
+				     __PPC_RA(base) | ((i) & 0xfffc))
+
+#define PPC_LD(r, base, i)	EMIT(PPC_INST_LD | __PPC_RT(r) |	      \
+				     __PPC_RA(base) | IMM_L(i))
+#define PPC_LWZ(r, base, i)	EMIT(PPC_INST_LWZ | __PPC_RT(r) |	      \
+				     __PPC_RA(base) | IMM_L(i))
+#define PPC_LHZ(r, base, i)	EMIT(PPC_INST_LHZ | __PPC_RT(r) |	      \
+				     __PPC_RA(base) | IMM_L(i))
+/* Convenience helpers for the above with 'far' offsets: */
+#define PPC_LD_OFFS(r, base, i) do { if ((i) < 32768) PPC_LD(r, base, i);     \
+		else {	PPC_ADDIS(r, base, IMM_HA(i));			      \
+			PPC_LD(r, r, IMM_L(i)); } } while(0)
+
+#define PPC_LWZ_OFFS(r, base, i) do { if ((i) < 32768) PPC_LWZ(r, base, i);   \
+		else {	PPC_ADDIS(r, base, IMM_HA(i));			      \
+			PPC_LWZ(r, r, IMM_L(i)); } } while(0)
+
+#define PPC_LHZ_OFFS(r, base, i) do { if ((i) < 32768) PPC_LHZ(r, base, i);   \
+		else {	PPC_ADDIS(r, base, IMM_HA(i));			      \
+			PPC_LHZ(r, r, IMM_L(i)); } } while(0)
+
+#define PPC_CMPWI(a, i)		EMIT(PPC_INST_CMPWI | __PPC_RA(a) | IMM_L(i))
+#define PPC_CMPDI(a, i)		EMIT(PPC_INST_CMPDI | __PPC_RA(a) | IMM_L(i))
+#define PPC_CMPLWI(a, i)	EMIT(PPC_INST_CMPLWI | __PPC_RA(a) | IMM_L(i))
+#define PPC_CMPLW(a, b)		EMIT(PPC_INST_CMPLW | __PPC_RA(a) | __PPC_RB(b))
+
+#define PPC_SUB(d, a, b)	EMIT(PPC_INST_SUB | __PPC_RT(d) |	      \
+				     __PPC_RB(a) | __PPC_RA(b))
+#define PPC_ADD(d, a, b)	EMIT(PPC_INST_ADD | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | __PPC_RB(b))
+#define PPC_MUL(d, a, b)	EMIT(PPC_INST_MULLW | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | __PPC_RB(b))
+#define PPC_MULHWU(d, a, b)	EMIT(PPC_INST_MULHWU | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | __PPC_RB(b))
+#define PPC_MULI(d, a, i)	EMIT(PPC_INST_MULLI | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | IMM_L(i))
+#define PPC_DIVWU(d, a, b)	EMIT(PPC_INST_DIVWU | __PPC_RT(d) |	      \
+				     __PPC_RA(a) | __PPC_RB(b))
+#define PPC_AND(d, a, b)	EMIT(PPC_INST_AND | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_RB(b))
+#define PPC_ANDI(d, a, i)	EMIT(PPC_INST_ANDI | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | IMM_L(i))
+#define PPC_AND_DOT(d, a, b)	EMIT(PPC_INST_ANDDOT | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_RB(b))
+#define PPC_OR(d, a, b)		EMIT(PPC_INST_OR | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_RB(b))
+#define PPC_ORI(d, a, i)	EMIT(PPC_INST_ORI | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | IMM_L(i))
+#define PPC_ORIS(d, a, i)	EMIT(PPC_INST_ORIS | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | IMM_L(i))
+#define PPC_SLW(d, a, s)	EMIT(PPC_INST_SLW | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_RB(s))
+#define PPC_SRW(d, a, s)	EMIT(PPC_INST_SRW | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_RB(s))
+/* slwi = rlwinm Rx, Ry, n, 0, 31-n */
+#define PPC_SLWI(d, a, i)	EMIT(PPC_INST_RLWINM | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_SH(i) |	      \
+				     __PPC_MB(0) | __PPC_ME(31-(i)))
+/* srwi = rlwinm Rx, Ry, 32-n, n, 31 */
+#define PPC_SRWI(d, a, i)	EMIT(PPC_INST_RLWINM | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_SH(32-(i)) |	      \
+				     __PPC_MB(i) | __PPC_ME(31))
+/* sldi = rldicr Rx, Ry, n, 63-n */
+#define PPC_SLDI(d, a, i)	EMIT(PPC_INST_RLDICR | __PPC_RA(d) |	      \
+				     __PPC_RS(a) | __PPC_SH(i) |	      \
+				     __PPC_MB(63-(i)) | (((i) & 0x20) >> 4))
+#define PPC_NEG(d, a)		EMIT(PPC_INST_NEG | __PPC_RT(d) | __PPC_RA(a))
+
+/* Long jump; (unconditional 'branch') */
+#define PPC_JMP(dest)		EMIT(PPC_INST_BRANCH |			      \
+				     (((dest) - (ctx->idx * 4)) & 0x03fffffc))
+/* "cond" here covers BO:BI fields. */
+#define PPC_BCC_SHORT(cond, dest)	EMIT(PPC_INST_BRANCH_COND |	      \
+					     (((cond) & 0x3ff) << 16) |	      \
+					     (((dest) - (ctx->idx * 4)) &     \
+					      0xfffc))
+#define PPC_LI32(d, i)		do { PPC_LI(d, IMM_L(i));		      \
+		if ((u32)(uintptr_t)(i) >= 32768) {			      \
+			PPC_ADDIS(d, d, IMM_HA(i));			      \
+		} } while(0)
+#define PPC_LI64(d, i)		do {					      \
+		if (!((uintptr_t)(i) & 0xffffffff00000000ULL))		      \
+			PPC_LI32(d, i);					      \
+		else {							      \
+			PPC_LIS(d, ((uintptr_t)(i) >> 48));		      \
+			if ((uintptr_t)(i) & 0x0000ffff00000000ULL)	      \
+				PPC_ORI(d, d,				      \
+					((uintptr_t)(i) >> 32) & 0xffff);     \
+			PPC_SLDI(d, d, 32);				      \
+			if ((uintptr_t)(i) & 0x00000000ffff0000ULL)	      \
+				PPC_ORIS(d, d,				      \
+					 ((uintptr_t)(i) >> 16) & 0xffff);    \
+			if ((uintptr_t)(i) & 0x000000000000ffffULL)	      \
+				PPC_ORI(d, d, (uintptr_t)(i) & 0xffff);	      \
+		} } while (0);
+
+static inline bool is_nearbranch(int offset)
+{
+	return (offset < 32768) && (offset >= -32768);
+}
+
+/*
+ * The fly in the ointment of code size changing from pass to pass is
+ * avoided by padding the short branch case with a NOP.	 If code size differs
+ * with different branch reaches we will have the issue of code moving from
+ * one pass to the next and will need a few passes to converge on a stable
+ * state.
+ */
+#define PPC_BCC(cond, dest)	do {					      \
+		if (is_nearbranch((dest) - (ctx->idx * 4))) {		      \
+			PPC_BCC_SHORT(cond, dest);			      \
+			PPC_NOP();					      \
+		} else {						      \
+			/* Flip the 'T or F' bit to invert comparison */      \
+			PPC_BCC_SHORT(cond ^ COND_CMP_TRUE, (ctx->idx+2)*4);  \
+			PPC_JMP(dest);					      \
+		} } while(0)
+
+/* To create a branch condition, select a bit of cr0... */
+#define CR0_LT		0
+#define CR0_GT		1
+#define CR0_EQ		2
+/* ...and modify BO[3] */
+#define COND_CMP_TRUE	0x100
+#define COND_CMP_FALSE	0x000
+/* Together, they make all required comparisons: */
+#define COND_GT		(CR0_GT | COND_CMP_TRUE)
+#define COND_GE		(CR0_LT | COND_CMP_FALSE)
+#define COND_EQ		(CR0_EQ | COND_CMP_TRUE)
+#define COND_NE		(CR0_EQ | COND_CMP_FALSE)
+#define COND_LT		(CR0_LT | COND_CMP_TRUE)
+
+#define SEEN_DATAREF 0x10000 /* might call external helpers */
+#define SEEN_XREG    0x20000 /* X reg is used */
+#define SEEN_MEM     0x40000 /* SEEN_MEM+(1<<n) = use mem[n] for temporary
+			      * storage */
+#define SEEN_MEM_MSK 0x0ffff
+
+struct codegen_context {
+	unsigned int seen;
+	unsigned int idx;
+	int pc_ret0; /* bpf index of first RET #0 instruction (if any) */
+};
+
+#endif
+
+#endif

arch/powerpc/net/bpf_jit_64.S

+/* bpf_jit.S: Packet/header access helper functions
+ * for PPC64 BPF compiler.
+ *
+ * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+
+#include <asm/ppc_asm.h>
+#include "bpf_jit.h"
+
+/*
+ * All of these routines are called directly from generated code,
+ * whose register usage is:
+ *
+ * r3		skb
+ * r4,r5	A,X
+ * r6		*** address parameter to helper ***
+ * r7-r10	scratch
+ * r14		skb->data
+ * r15		skb headlen
+ * r16-31	M[]
+ */
+
+/*
+ * To consider: These helpers are so small it could be better to just
+ * generate them inline.  Inline code can do the simple headlen check
+ * then branch directly to slow_path_XXX if required.  (In fact, could
+ * load a spare GPR with the address of slow_path_generic and pass size
+ * as an argument, making the call site a mtlr, li and bllr.)
+ *
+ * Technically, the "is addr < 0" check is unnecessary & slowing down
+ * the ABS path, as it's statically checked on generation.
+ */
+	.globl	sk_load_word
+sk_load_word:
+	cmpdi	r_addr, 0
+	blt	bpf_error
+	/* Are we accessing past headlen? */
+	subi	r_scratch1, r_HL, 4
+	cmpd	r_scratch1, r_addr
+	blt	bpf_slow_path_word
+	/* Nope, just hitting the header.  cr0 here is eq or gt! */
+	lwzx	r_A, r_D, r_addr
+	/* When big endian we don't need to byteswap. */
+	blr	/* Return success, cr0 != LT */
+
+	.globl	sk_load_half
+sk_load_half:
+	cmpdi	r_addr, 0
+	blt	bpf_error
+	subi	r_scratch1, r_HL, 2
+	cmpd	r_scratch1, r_addr
+	blt	bpf_slow_path_half
+	lhzx	r_A, r_D, r_addr
+	blr
+
+	.globl	sk_load_byte
+sk_load_byte:
+	cmpdi	r_addr, 0
+	blt	bpf_error
+	cmpd	r_HL, r_addr
+	ble	bpf_slow_path_byte
+	lbzx	r_A, r_D, r_addr
+	blr
+
+/*
+ * BPF_S_LDX_B_MSH: ldxb  4*([offset]&0xf)
+ * r_addr is the offset value, already known positive
+ */
+	.globl sk_load_byte_msh
+sk_load_byte_msh:
+	cmpd	r_HL, r_addr
+	ble	bpf_slow_path_byte_msh
+	lbzx	r_X, r_D, r_addr
+	rlwinm	r_X, r_X, 2, 32-4-2, 31-2
+	blr
+
+bpf_error:
+	/* Entered with cr0 = lt */
+	li	r3, 0
+	/* Generated code will 'blt epilogue', returning 0. */
+	blr
+
+/* Call out to skb_copy_bits:
+ * We'll need to back up our volatile regs first; we have
+ * local variable space at r1+(BPF_PPC_STACK_BASIC).
+ * Allocate a new stack frame here to remain ABI-compliant in
+ * stashing LR.
+ */
+#define bpf_slow_path_common(SIZE)				\
+	mflr	r0;						\
+	std	r0, 16(r1);					\
+	/* R3 goes in parameter space of caller's frame */	\
+	std	r_skb, (BPF_PPC_STACKFRAME+48)(r1);		\
+	std	r_A, (BPF_PPC_STACK_BASIC+(0*8))(r1);		\
+	std	r_X, (BPF_PPC_STACK_BASIC+(1*8))(r1);		\
+	addi	r5, r1, BPF_PPC_STACK_BASIC+(2*8);		\
+	stdu	r1, -BPF_PPC_SLOWPATH_FRAME(r1);		\
+	/* R3 = r_skb, as passed */				\
+	mr	r4, r_addr;					\
+	li	r6, SIZE;					\
+	bl	skb_copy_bits;					\
+	/* R3 = 0 on success */					\
+	addi	r1, r1, BPF_PPC_SLOWPATH_FRAME;			\
+	ld	r0, 16(r1);					\
+	ld	r_A, (BPF_PPC_STACK_BASIC+(0*8))(r1);		\
+	ld	r_X, (BPF_PPC_STACK_BASIC+(1*8))(r1);		\
+	mtlr	r0;						\
+	cmpdi	r3, 0;						\
+	blt	bpf_error;	/* cr0 = LT */			\
+	ld	r_skb, (BPF_PPC_STACKFRAME+48)(r1);		\
+	/* Great success! */
+
+bpf_slow_path_word:
+	bpf_slow_path_common(4)
+	/* Data value is on stack, and cr0 != LT */
+	lwz	r_A, BPF_PPC_STACK_BASIC+(2*8)(r1)
+	blr
+
+bpf_slow_path_half:
+	bpf_slow_path_common(2)
+	lhz	r_A, BPF_PPC_STACK_BASIC+(2*8)(r1)
+	blr
+
+bpf_slow_path_byte:
+	bpf_slow_path_common(1)
+	lbz	r_A, BPF_PPC_STACK_BASIC+(2*8)(r1)
+	blr
+
+bpf_slow_path_byte_msh:
+	bpf_slow_path_common(1)
+	lbz	r_X, BPF_PPC_STACK_BASIC+(2*8)(r1)
+	rlwinm	r_X, r_X, 2, 32-4-2, 31-2
+	blr